// CHECK-NEXT:      %18 = "tfl.pseudo_const"() <{value = dense<2> : tensor<1xi32>}> : () -> tensor<1xi32>
// CHECK-NEXT:      %19 = "tfl.concatenation"(%18, %17, %16) <{axis = 0 : i32, fused_activation_function = "NONE"}> : (tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<3xi32>
// CHECK-NEXT:      %20 = mhlo.dynamic_reshape %2, %19 : (tensor<2x3x4x?xf32>, tensor<3xi32>) -> tensor<2x3x?xf32>

    } : (tensor<10xf32>, tensor<10xf32>, tensor<6xi32>) -> (tensor<10xf32>, tensor<6xi32>)
    // CHECK: return %[[RESULT_CAST]], %[[RESULT]]#1 : tensor<10xf32>, tensor<6xi32>
    func.return %0#0, %0#1 : tensor<10xf32>, tensor<6xi32>
  }

  // CHECK-LABEL: func private @main_0
  // CHECK-SAME: %[[ARG_0:.*]]: tensor<10xbf16>, %[[ARG_1:.*]]: tensor<10xbf16>, %[[ARG_2:.*]]: tensor<6xi32>
  func.func private @main_0(

    %2 = stablehlo.reshape %1 : (tensor<i32>) -> tensor<1xi32>
    %3 = stablehlo.concatenate %2, %cst_0, %cst_1, %cst_2, dim = 0 : (tensor<1xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<4xi32>
    %4 = stablehlo.dynamic_broadcast_in_dim %arg2, %3, dims = [0, 1, 2, 3] : (tensor<1x1x1x2xf32>, tensor<4xi32>) -> tensor<?x3x4x2xf32>
    %5 = stablehlo.add %0, %4 : tensor<?x3x4x2xf32>

      %17 = "tfl.reduce_prod"(%16, %cst_14) {keep_dims = true, tac.device = "CPU", tac.inference_type = "FLOAT"} : (tensor<2xi32>, tensor<1xi32>) -> tensor<1xi32>
      %18 = "tfl.reshape"(%arg1, %17) {tac.device = "DARWINN", tac.inference_type = "FLOAT"} : (tensor<?x?x!tf_type.string>, tensor<1xi32>) -> tensor<?x!tf_type.string>

    %4 = "tf.LayerNorm"(%a1, %a2, %a3, %a4) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>
   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

// CHECK-NEXT:  %[[argmax:.*]] = "tfl.arg_max"(%[[softmax]], %[[cst]]) : (tensor<128x16xf32>, tensor<i32>) -> tensor<128xi32>
  %3 = "tfl.arg_max"(%2, %cst) : (tensor<128x16xf32>, tensor<i32>) -> tensor<128xi32>
// CHECK-NEXT:  return %[[softmax]], %[[argmax]] : tensor<128x16xf32>, tensor<128xi32>
  func.return %2, %3 : tensor<128x16xf32>, tensor<128xi32>
}

// CHECK-LABEL: PruneUnusedLstm

    >} : (
      tensor<?x?x6xi8>, tensor<6x8xi8>
    ) -> tensor<?x?x8xi32>
  %1 = chlo.broadcast_add %0, %zp_offset : (
      tensor<?x?x8xi32>, tensor<8xi32>) -> tensor<?x?x8xi32>
  %2 = chlo.broadcast_add %1, %bias : (
      tensor<?x?x8xi32>, tensor<8xi32>) -> tensor<?x?x8xi32>
  return %2 : tensor<?x?x8xi32>
}

// -----

// CHECK-LABEL: func @dot_general_add_add_static

    %4 = "tf.LayerNorm"(%a1, %a2, %a3, %a4) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>
   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

  module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 268 : i32}} {
    func.func @main() -> tensor<1xi32> {
      %0 = "tf.Const"() {value = dense<1000> : tensor<1xi32>} : () -> tensor<1xi32>
      func.return %0 : tensor<1xi32>
    }
  })";

TEST(LegalizeMlirTest, LegalizesModule) {
  mlir::DialectRegistry mlir_registry;
  RegisterAllTensorFlowDialects(mlir_registry);

    %26 = "tfl.slice"(%arg3, %22, %25) : (tensor<*xf32>, tensor<?xi32>, tensor<?xi32>) -> tensor<*xf32>
    %27 = "tfl.reshape"(%5, %cst_12) : (tensor<i32>, tensor<1xi32>) -> tensor<1xi32>
    %28 = "tfl.concatenation"(%27, %21) {axis = 0 : i32, fused_activation_function = "NONE"} : (tensor<1xi32>, tensor<?xi32>) -> tensor<?xi32>